A map of the software-engineering process as practiced before AI coding agents became the norm — built cell-by-cell to contrast the then (pre-2022) against the now (post-2023 AI-agent era).
Cutoff: ≤ November 2022 · pre-ChatGPTScoped to web / cloud-era SDLC · snapshot as of Jul 2026
Then ≤ Nov 2022Now post-2023 (AI agents)Inflection: ChatGPT, Nov 30 2022Also spans this cell
A minor issues badge means a residual date or attribution nuance — shown in each node's Adversarial fact-check panel, not a factual hole.
The map — lifecycle phases × ecosystem layers
Filled cells are researched nodes (click for detail). Empty cells mean no dedicated node — the practice is covered by an adjacent cell or the contrast rail. Toggle After · agents to overlay the factory agent at every phase.
Contrast anchors — then vs now
Evolution — from craft to factory (1968 → 2026)
How much did AI agents disrupt each area?
Per-area disruption rated 0 (virtually unchanged) → 5 (radically transformed) during the sanitize pass, grounded in each node's before/after contrast. Illustrative expert ratings, not measurements.
By area
Where AI changed things most — by cluster
The "After" — software engineering becomes factory engineering
Distilled from Zach Lloyd (Warp), "Self-Improving Software Factories: The New Open Source Model" — AI Engineer World's Fair 2026. source talk. This is the "where it's heading" capstone that sharpens the contrast — the verified research nodes above are the "before"; this is the concrete vision of the "after."
"Software engineering will become factory engineering."
…and "You won't be building the product. You'll be building the thing that builds the product." The engineer's locus shifts from writing code to designing and operating the factory that writes code.
The factory loop — an agent at every stage of the SDLC
↻ Outer loop (self-improvement): a skill-updater agent reviews outcomes & comments, then opens PRs that improve the other agents' SKILL.mds — the factory compounds on itself. No pre-2022 analog exists for this.
Four pillars of an effective factory
Automation
Work moves itself through the pipeline — agents triage, spec, implement, review, and verify.
Context + Skills
Reusable, retrievable skills + RAG give agents the repo's knowledge, conventions, and tool access.
Human-in-the-loop
Humans review specs, code, and verification output — now the rate-limiting, highest-leverage step.
Self-improvement
Inner loop ships software; outer loop upgrades the skills. The factory gets better at building itself.
The new unit of measurement
Then · ≤ 2022
DORA & SPACE measured human throughput — deploy frequency, lead time, MTTR, change-failure rate; plus activity, communication, efficiency & satisfaction.
Now · factory era
Factory efficiency = Software shipped ÷ (token cost + human cost)A new denominator appears: AI token spend joins human time on the bottom line.